Annotating If the Authors of a Tweet are Located at the Locations They Tweet About

Annotating If Authors of Tweets are Located in the Locations They Tweet About Vivek Doudagiri, Alakananda Vempala and Eduardo Blanco Human Intelligence and Language Technologies Lab University of North Texas VivekReddyDoudagiri@my.unt.edu, AlakanandaVempala@my.unt.edu, eduardo.blanco@unt.edu Abstract The locations in a tweet do not always indicate spatial information involving the author of the tweet. In this paper, we investigate whether authors are located or not located in the locations they tweet about, and temporally anchor this spatial information in the tweet timestamp. Specifically, we work with temporal tags centered around the tweet timestamp: longer than 24 hours before or after tweeting, within 24 hours before or after tweeting, and at the time of tweeting. We introduce a corpus of 1,200 location mentions from 1,062 tweets, discuss several annotation samples, and analyze annotator disagreements. Keywords: Spatial knowledge, Social Networks, Twitter 1. Introduction Twitter has quickly become one of the most popular so- cial media sites. It has 313 million monthly active users, and 500 million tweets are published daily. People tweet about their current locations (e.g., I’m at Walgreens in Ana- heim, California), as well as past (e.g., I miss living in Ker- man) and (probable) future locations (e.g., Can’t wait to visit Italy!). Tweets often contain hints regarding how long the author is in a particular location, either implicitly (e.g., people usually stay at pharmacies such as Walgreens for a few minutes to an hour, not days) or explicitly (e.g., My 2 day vacation to San Diego beach starts tomorrow). In this work, we present a corpus annotating whether the author of a tweet is located in the locations mentioned in his tweets. Going beyond named entity recognition, we an- notate whether the author is located or not located in the lo- cation he tweets about with respect to the time he tweeted (before, during and after). To the best of our knowledge, this problem has not been explored before. We found that no spatial relationship can be inferred between authors of tweets and the locations they tweet about in 21% of in- stances. In other words, 1 out of 5 locations in a tweet do not indicate any spatial information about the author. The major contributions of this paper are: 1. We create a corpus of 1,062 tweets containing 1,200 location named entities, and annotate whether the au- thors are located or not located in those locations with respect to the time they tweeted (when the author tweeted, within 24 hours before and after he tweeted, and longer than 24 hours before and after he tweeted).1 2. We provide several annotation examples and the label distributions per temporal tag. 3. We present detailed inter-annotator agreement calcu- lations, including Cohen’s  and confusion matrices per temporal tag. 1Available at https://alakanandav.bitbucket. io/ 2. Previous Work The work presented here is inspired by Li and Sun (2014), who work with points of interest in tweets using the Foursquare API (their points of interest are similar to our locations). Li and Sun determine if the author …

Content truncated to 500 words